AI tests old ideas about navigation and the brain-and comes up empty in young adults
Steven Weisberg at The University of Texas at Arlington reports a clear null: advanced AI models could not find a reliable link between brain structure and navigation ability in healthy young adults. The work challenges the long-standing assumption that better wayfinding maps cleanly onto bigger or differently shaped brain regions.
Alongside University of Florida Ph.D. candidate Ashish Sahoo, the team pushed beyond simple volumetrics using deep convolutional neural networks and other machine-learning approaches on MRI data. Even with these tools, structure-based features failed to predict who learned and recalled routes better.
Why this matters
Navigation is tied to independence, memory, and dementia risk. A null result here signals that, at least in this age range and with current MRI structural measures, the effect (if it exists) is small or buried under noise. For researchers, it's a cue to rethink where the signal might live: task design, functional dynamics, connectivity, or longitudinal change-rather than static anatomy alone.
Methods in brief
N=90 healthy adults (mean age ≈ 23.1) learned two routes in a virtual environment. The analyses contrasted the hippocampus-a usual suspect for spatial memory-with a control region (thalamus). Models built on hippocampal structure did not outperform those built on the control region, and none met thresholds you'd want to stake a claim on.
The study was conducted at the University of Florida before Weisberg joined UT Arlington last fall as part of the RISE 100 initiative. It is published in Neuropsychologia (journal page).
Key takeaways you can apply
- Interrogate measurement first. Use high-reliability navigation tasks and report test-retest where possible.
- Right-size the sample for ML. Deep models on MRI need big-N or multi-site data to beat simple baselines.
- Move beyond structure-only features. Add functional signals (task fMRI, resting connectivity), diffusion metrics, and behavior-rich logs.
- Lock down validation. Hold out participants, check leakage, and benchmark against transparent linear models.
- Estimate the noise ceiling. If behavior reliability caps variance explained, don't expect miracles from model complexity.
- Pre-register analysis plans and share code to make nulls informative and replicable.
What the researchers say
"With the quality of data we have from MRI scans and this healthy young adult population, there does not appear to be a detectable signal using these advanced metrics," Weisberg said.
He added that AI has been effective at predicting disease states, and the open question is whether these models will prove useful for everyday behavioral functions such as training and education.
Where this goes next
The team points to larger samples and older populations as the next step. If structure-behavior links emerge anywhere, they may show up with age-related change, clinical risk, or more sensitive behavioral assays.
Big picture: the absence of a clean structural marker in young adults doesn't diminish AI's value-it sharpens the research agenda. Expect more multi-modal datasets, better task reliability, and models that test clear, falsifiable hypotheses about how brain features map to behavior.
For researchers building AI skills
If you want a structured way to level up the exact methods used in studies like this-data pipelines, modeling, validation-see the AI Learning Path for Research Scientists.
Your membership also unlocks:
